A Learning-based Fast Image Super-resolution And Its Implementation

With the advent of the digital age,display devices supporting ultra-high resolution such as 4k or even 8k are gradually appearing in people's daily lives,and the demand for high-resolution,high-definition multimedia services is increasing.Limited to expensive acquisition devices,limited data bandwidth,and scarce Ultra HD video resources,high resolution images or video are often not readily available.Image super-resolution technology can reconstruct images with higher resolution than the imaging device in a low-cost way.It is an important research direction in the field of image processing and has high practical value.In order to reconstruct high-quality super-resolution images,a low-complexity and high-performance image super-resolution algorithm is particularly important.The fast super-resolution algorithm based on edge clustering(SI algorithm)is one of the sample learning methods,which can achieve single image super-resolution with magnification of 2.This algorithm directly inputs the linear mapping relationship between high and low resolution image blocks.The low-resolution image mapped to a high-resolution image is a fast and efficient image super-resolution algorithm,which does not require overlapping and averaging operations.This thesis is based on the SI algorithm.The main research work is as follows:(1)The SI algorithm simply clusters the training samples based on the edge gradients in the horizontal and vertical directions of the image,and the clustering results are rough.In order to solve this problem,this thesis proposes an improved algorithm.Based on the original algorithm,the improved algorithm designs a new edge feature extraction method and extracts more abundant image edge information.For the image feature vector,the improved algorithm adopts an effective dimensionality reduction strategy combined with the K-means clustering method to make the clustering result more accurate.Simulation experiments show that compared with SI algorithm,the improved algorithm effectively improves the quality of reconstructed images,has better performance on subjective and objective evaluation indicators,and maintains the low computational complexity of the original algorithm,suitable for low hardware.Power consumption is achieved.(2)Perform RTL-level design and simulation on the improved algorithm,and analyze the implementation process of each module of the algorithm.By comparing the hardware simulation results of the improved algorithm with the software implementation results,the accuracy and reliability of the hardware simulation are verified.(3)In order to study the implementation of SI algorithm in hardware system,a small HD camera system and a 4k ultra-high definition camera system were designed.Firstly,a small HD camera system was designed and implemented.The overall structure and key device circuit of the camera system were introduced.Secondly,the SI algorithm was applied to the small camera system.Based on this system,the video resolution of 800×600@60Hz was improved,and a super-resolution video image of 1600x1200@60Hz was generated.Finally,the improved algorithm is applied to the 4k Ultra HD camera system,providing a solution that can support 4k@30Hz video output.